People say out loud what they need. We hear it — in a minute.

Sonar watches public Telegram chats and finds the people who are looking for what you sell right now. A language model tells a live request apart from small talk, ads and competitors, and the alert lands in your Telegram within a minute.

Try it free for 30 days Open in Telegram See it for my niche No card. From $49 a month afterwards — and if there is no volume in your niche, we say so before you pay.

< 60 secfrom the message to the alert

90–95%cut before the model — the economics

Read-onlythe accounts never write to chats

Niche = configa new market without touching the code

How it works

Four filters between the firehose and your alert

Two hundred sources produce tens of thousands of messages a day. Sending all of them to a model would break the unit economics before the first customer — so the cheap layers do the cutting and the model only judges what survives.

1 · Collector Read-only userbots receive every message from watched groups in real time.

2 · Pre-filter Language, niche dictionaries, ad patterns, author blacklist. No model, no cost.

3 · The model One structured verdict: is it a lead, who is the author, budget, geo, urgency — and why.

4 · Dedup + enrich The same person asking in five chats is one lead, with their history attached.

5 · Alert A card in Telegram with one-tap feedback that tunes the filter.

Every message is kept — including the ones the pre-filter cut. Without them there is no way to measure what the filter ate.

Niches

One engine, different businesses

A niche is a config: what counts as a lead, which words hint at it, how sure the model must be. Adding a market means adding a niche, not writing code. A single chat can feed several niches at once.

niche · turkey_rent

Property rental, Turkey

A person looks for a home for themselves or their family. Not an agent, not an owner offering a unit.

looking forneed a flatany recommendationsmoving to

stop words: for rent · for sale · nightly

niche · llm_inference

Cheap LLM inference

Someone buying inference for their own product: token prices, a cheaper API, access to specific models.

cheaper than openaiprice per tokenneed an apigpu rental

stop words: selling api access · promo code

niche · dev_consulting

Paid development and consulting

Someone ready to pay to get a stuck product shipped: fix what was vibe-coded, sort out the deploy.

who can take this onhappy to payneed a dev urgentlydeploy is broken

stop words: looking for work · offering my services

Not a black box

Every decision is written down — including the rejections

The hardest call in this market: half the “seekers” in a rental chat are agents hunting for clients. The model states the author's role explicitly and the alert only fires on a real client. Every verdict is stored with its reason, so a filter that misses can be debugged instead of guessed at.

verdict log · demo data

Verdict	Message	Why	Confidence
lead client	“Looking for a 2+1 in Konyaaltı up to 30k, for a year, from September”	Direct request: district, budget and timeframe named	86%
not a lead agent	“Looking for a 3+1 for a client, budget negotiable, urgent”	An agent searching on behalf of a client — a competitor, not a lead	81%
not a lead competitor	“Renting out a cosy studio in Mahmutlar, 15k/mo, DM me”	An offer, not a request	93%
not a lead unclear	“Anyone know a good dentist in Antalya?”	Off-niche: a request, but not for what we sell	95%

Rejection reasons are categorised — spam, ad, keyword coincidence, wrong geo, competitor — which is what makes tuning the prompt possible.

What the customer gets

A card you can act on, and a filter that learns from you

🔥 Lead · Rental · Antalya

Budget: ~30 000 TRY/mo (stated)
Urgency: high · from September
Author: @username · first message

“Looking for a 2+1 in Konyaaltı up to 30k, for a year, from September. Close to the sea if possible”

🔗 Open ✅ Take it ❌ Not it 🚫 Spam

Demo card. One tap of feedback recalibrates thresholds, blacklists spammers and feeds the regression dataset.

Precision over volume The target is 80% precision and 70% recall, measured weekly on a hand-read sample — not a stream of “maybe” alerts nobody opens.

The filter is measurable Pass-through share, rejection reasons, per-source yield: a chat with traffic and no leads is visible and gets switched off.

Safe by construction Accounts only read, join rates are capped, sessions are pinned to their proxy and geo. Zero bans is a product metric.

Prompt changes are scored Every edit runs against a labelled dataset, tagged with its prompt version — quality moves on purpose, not by luck.

All examples on this page are synthetic. Messages, authors and chats from monitored sources are never published — they stay inside the operator's panel.

Start with your own niche

Tell the bot what you sell and where. We will find the chats, measure them, and show you how many matching requests they produced last week — before you pay anything.

Start free Open the bot Ask a question

Sonar · Mingles AI LTD Bot Niches Cases · Cases Docs FAQ Support Data Terms Operator sign-in